Reflection AI unveils Beam, a 501B-parameter open model due this month

Key Summary

Reflection AI has unveiled Beam, an open-weight AI model with 501 billion total parameters, expected to be released later this month. The model, based on a sparse Mixture-of-Experts architecture, uses only 23 billion active parameters, significantly reducing inference compute costs. Beam was pretrained on 23.8 trillion tokens and underwent large-scale reinforcement learning on 10.5K NVIDIA GPUs. It is positioned against Chinese rivals GLM-5.2 and Qwen 3.8-Max, matching or surpassing performance in reasoning and coding. Beam achieved a score of 80.9 on SWE-Bench Verified and 80.1 on Terminal Bench v2.1, with an efficiency advantage of 3-4 times less inference compute.